The third in a series of military AI summits was held in La Coruña, Spain in February 2026. The aim of the meeting was to convert previously agreed principles on the military use of AI into action. The summit was attended by government officials, military personnel, representatives from industry and researchers from thinktanks.
The goal of many experts and policymakers in this area is to usher countries towards a regulatory framework on using machine intelligence in warfare. To this end, the latest Responsible AI in the Military Domain (REAIM) summit presented a non-binding commitment for countries to sign.
The REAIM agreement affirmed the need for human oversight of military AI systems, called for countries to carry out risk assessments and robust testing, and committed to transparency on how decisions are made when using AI in conflicts.
The reasoning behind such recommendations is sound. However, translating such a framework from plan to action faces multiple hurdles. Ultimately, less than half of the countries represented at this year’s REAIM summit signed the non-binding commitment.
To understand why, it’s instructive to look at what happened at the 80th UN General Assembly held in New York in December 2025. At the meeting, members of the assembly’s first committee voted overwhelmingly to approve two resolutions calling for greater international scrutiny of the risks from military uses of AI. However, the US and Russia notably opposed the resolutions.
The US had been a signatory to earlier REAIM summit commitments. But this year, the US and China both declined to sign it. There seems little doubt that this helped fuel the hesitancy of other countries.
The Netherlands’ defence minister Ruben Brekelmans put it succinctly when he said that governments face a “prisoner’s dilemma”. This is a concept in game theory where two rational individuals face competing incentives to cooperate with or betray one other.
Countries are effectively having to implement responsible restrictions on military AI without subjecting their armed forces to limitations that could be exploited by a less conscientious enemy.
Devdroid, CC BY-SA
An important sticking point is the deployment of autonomous AI systems in warfare. The idea of autonomous weapons systems, which make decisions without input from a human, remains a grave concern for many interested parties on this issue.
There continues to be a consensus against using such weapons. But countries can’t reach a common position over how to define them, particularly so-called lethal autonomous weapons systems – or Laws for short. These are often characterised as “killer robots”, though a more detailed description remains elusive.
A uniform definition for such systems could be an important first step towards a discussion on regulation. But, despite efforts by academic experts to draft and amend flexible definitions, countries remain too far apart on the characteristics they ascribe to these weapons.
The impasse is informed by a fear that accepting a definition could restrict countries’ militaries on the battlefield – threatening national security.
Testing grounds for tech
Existing legal mechanisms, such as international humanitarian laws, already prohibit the irresponsible and unethical use of military AI – in theory, at least. But how these laws would function in practice when applied to real world scenarios is uncertain.
The ongoing Russia-Ukraine war, the war in Gaza and the more recent escalation in Iran are being used by militaries as testing grounds for such technology.
The Lavender intelligence gathering and targeting software, used by Israel in Gaza, and Anthropic’s AI Model Claude, used by the US in Iran, demonstrate the rapid pace of advancement in AI-powered data gathering and analysis. This can help military planners make quicker decisions.
Drone warfare – AI assisted, autonomous and semi-autonomous – has grown at an equally rapid rate. This emerging technology is evolving significantly faster than the potential rules that could govern its use.
US Army / Staff Sgt Thomas Moeger
There’s a recurring argument that humans in the loop can operate as effective safeguards against the misuse of military AI systems. But as human overseers become familiar with the AI systems they use, their engagement may slip, causing them to become detached from the process.
As this happens, they may start to view real people as mere objects on a screen. This effect is known as automation bias. In such instances, human oversight could cease to be meaningful and instead lead to the simple rubber stamping of recommendations made by AI.
Additionally, the downsides of AI technology, such as bias, misinformation and disinformation generated by the systems themselves, and the erosion of human judgement resulting from overreliance on these systems, are not easy to solve after they enter use. This is why the REAIM summit commitment recommended risk assessments and robust testing before AI systems are adopted by militaries.
Without regulation, the risk of harm caused by AI systems remains significant. The severity of such risks balloons in magnitude when they are applied to military contexts. Miscalculations can lead to unintended escalation, as well as civilian deaths.
